Hacker News Reader

> The SSN name data seems almost certainly flawed to a small degree...I.e. It's just hard to believe that there are dozens of boys named Jennifer (and yet, strangely, no boys named Sue!)...but we're talking about infinitesimal rounding errors. The vast, vast majority of names are 99% one way or the other...with a few exceptions such as Leslie...though you can mitigate that by using older years of the SSN database.

I agree that the SSN data has flaws, but I only took names with at least 20 people. But the classification is probably iffy, as some names are classified as male and female.

> So I have to strongly disagree with OP that 80% accuracy is something to be astounded by when it comes to gender classification...

I originally hypothesized, I could reach 90% accuracy, but I could only get up to 82% max. As stated in the blog, 80% is the accuracy of a mammogram detecting cancer in a 40-45 year old woman which is pretty good for 3 features!

> I wonder how much more using Soundex would add to the accuracy? Creating a trained name classifier would be a fun project in service of a tool that could gender classify how masculine or feminine a made-up name sounds like...which would be a slightly useful tool if you were a fantasy fiction writer, though I suppose if you were to be a successful writer, your ear would be trained well enough for he purpose to not delegate it to a computational tool.

This would be very interesting to see!

I agree with your comment in the sense that 80% is not, as it seems to be, a very large number. One has to understand that in statistics the baseline technique is the "random". In the case of genders for USA, a random technique would probably have around 50% accuracy (http://kff.org/other/state-indicator/distribution-by-gender/), so 80% is just 30% more than the baseline (there's still the other 2/5 to cover).

I used to work with Text-Mining techniques and it was quite easy to come up with a statistical metric that could reach 70%-75% accuracy, for almost all problems that I've dealt with (mainly in the sub-area of concept extraction - information retrieval). I assume it is similar for this guy, so that is why he reaches this kind of values very fast. For instance, with only the character frequency and order he reports results of 75% which, if you look carefully, is just 20%-25% above the random baseline, not "pure" 75%..

Then he trains a classifier (Random Forest) and extracts the most relevant features that the classifier used, which is existence of an 'a' character in the last two positions of the name (the order of 'a' feature seems quite irrelevant to me). With these features he reaches the 80% ("random" + 30%) accuracy which is a good number but still far. I would say that these two features are to be expected, as, at least in my mother tongue (Portuguese), feminine names end with an 'a' (eg. Maria, Ana, Roberta [vs Roberto - masc], etc..).

But all in all it can be considered a good exercise!

Note that the statistics may depend on birth year for certain names.